How Sophisticated Are Neural Networks Needed to Predict Long-Term Nonadiabatic Dynamics?

Area of Science:

• Quantum dynamics
• Condensed phase chemistry
• Photochemistry

Background:

• Nonadiabatic dynamics are crucial for solar energy conversion and photochemical processes.
• Non-Markovian dynamics of the reduced density matrix in open quantum systems require knowledge of prior states for prediction.
• Understanding memory time is key to modeling these complex quantum systems.

Purpose of the Study:

• To explore time-series machine learning for predicting long-time nonadiabatic dynamics from short-time data.
• To compare machine learning methods with the physics-based Transfer Tensor Method (TTM).
• To investigate the impact of memory time on predictive accuracy and propose a method for estimating effective memory time.

Main Methods:

• Application of time-series machine learning models: Fully Connected Neural Network (FCN), Gated Recurrent Unit (GRU), and Convolutional Neural Network/Long Short-Term Memory (CNN-LSTM).
• Utilizing the Nakajima-Zwanzig generalized quantum master equation framework to represent non-Markovian dynamics as a linear map.
• Testing models on spin-boson, multistate harmonic (MSH) models, and a carotenoid-porphyrin-fullerene triad system.

Main Results:

• FCN models with linear mapping outperformed nonlinear models (GRU, CNN-LSTM) for systems with short memory times (spin-boson, MSH).
• Nonlinear CNN-LSTM and GRU models showed higher accuracy for triad MSH systems with long memory times.
• A practical method was developed to estimate effective memory time within a given tolerance.

Conclusions:

• The choice of machine learning model (linear vs. nonlinear) depends critically on the effective memory time of the quantum system.
• Time-series machine learning offers a powerful approach for predicting non-Markovian quantum dynamics.
• Findings provide guidance for applying ML to complex chemical and physical systems, particularly in energy conversion and photochemistry.